GitHub for a reinforcement learning research project consisting in simulating a novel Random Number Generator (RNG) by Deep Reinforcement Learning and Recurrent Neural Networks.

A PRNG is an algorithm generating pseudo-random numbers and in this research project it is approximated by a Deep Recurrent Neural Network using Reinforcement Learning. The network is trained to "randomly" generate a novel algorithm by Reinforcement Learning, using a deep agent to solve a partially observable Markov Decision Process. The formulation decouples the size of the action space and the length of the sequence, improving by a wide margin previous work's approach. Starting from a seed state (usually of zeros) the agent learns how to "move" in the highly dimensional environment by appending a sequence of bits with fixed length N to its current sequence, in order to reach states with high rewards. The reward is given by the result of the NIST test battery on the entire sequence computed at the last time step. The agent can only observe the last appended sequence at each time step and uses a RNN, specifically a LSTM, to memorize previous appended sequences, to devise an appropriate policy.

Link to previous work on arXiv. Link to previous work published article.

Link to the published article. Link to the arXiv article.

The algorithm used is Proximal Policy Optimization (PPO) with rewards-to-go and Generalized Advantage Estimation (GAE-Lambda) buffer.

License

The same of the article.

Framework used

• To run the NIST test battery NistRng (nistrng package) python implementation framework is used.
• To execute reinforcement learning the framework USienaRL (usienarl package) is used.

Compatible with Usienarl v0.7.6

Backend

• Python 3.6
• Tensorflow 1.15